Speaker: Dr Mang Chen
Date: Wednesday 17 February 2021
Time: 12:00 (noon)
Venue: Zoom (Register online)
The number of Earth observation satellites has increased drastically over the past decade, where some of these satellites enable the capture of two (or more) images of the same region at quasi real-time. Stereo vision techniques can then be used to automatically compute Digital Elevation Models (DEMs) that are important for a number of domains including hydrology, urban planning and natural hazard detection. These stereoscopically derived DEMs provide an efficient and low-cost means for remote mapping of surface topography over large areas and at multiple times for change detection.
The Centre National d’Etudes Spatiales (CNES) have developed the Stereo Pipeline for Pushbroom Images (S2P) framework that combines the information obtained from the Satellite together with a stereo matching process to estimate the DEM. However, the stereo matching process adopted by this framework is based on classical techniques.
The aim of the SAtellite TraIning and NETworking (SATINET) project is to adopt deep-learning based techniques to improve the stereo-matching process. WorldView-3 satellite images at a resolution 30cm and airborne Lidar data covering the area of San Fernando in Argentina was adopted in our evaluation. Compared with the inherent shortages of classical techniques in feature extraction for textureless, repeated pattern and occlusion, deep learning methods automatically learn and calculate feature parameters through training. Compared with the 66.85% completeness of the classical techniques (SGBM), our method reaches 74.05%, which is a 7% gain. The results below further show that our approach is more robust when compared to the state-of-the-art method.